Re: More filesystem need this fix (xfs: use MMAPLOCK around filemap_map_pages())

From: Hugh Dickins
Date: Thu Sep 17 2020 - 03:47:40 EST

On Thu, 17 Sep 2020, Dave Chinner wrote:
> On Wed, Sep 16, 2020 at 07:04:46PM -0700, Hugh Dickins wrote:
> > On Thu, 17 Sep 2020, Dave Chinner wrote:
> > > <pte now points to a freed page>
> >
> > No. filemap_map_pages() checks page->mapping after trylock_page(),
> > before setting up the pte; and truncate_cleanup_page() does a one-page
> > unmap_mapping_range() if page_mapped(), while holding page lock.
> Ok, fair, I missed that.
> So why does truncate_pagecache() talk about fault races and require
> a second unmap range after the invalidation "for correctness" if
> this sort of race cannot happen?

I thought the comment
* unmap_mapping_range is called twice, first simply for
* efficiency so that truncate_inode_pages does fewer
* single-page unmaps. However after this first call, and
* before truncate_inode_pages finishes, it is possible for
* private pages to be COWed, which remain after
* truncate_inode_pages finishes, hence the second
* unmap_mapping_range call must be made for correctness.
explains it fairly well. It's because POSIX demanded that when a file
is truncated, the user will get SIGBUS on trying to access even the
COWed pages beyond EOF in a MAP_PRIVATE mapping. Page lock on the
cache page does not serialize the pages COWed from it very well.

But there's no such SIGBUS requirement in the case of hole-punching,
and trying to unmap those pages racily instantiated just after the
punching cursor passed, would probably do more harm than good.

> Why is that different to truncate_pagecache_range() which -doesn't-i
> do that second removal? It's called for more than just hole_punch -
> from the filesystem's persepective holepunch should do exactly the
> same as truncate to the page cache, and for things like
> COLLAPSE_RANGE it is absolutely essential because the data in that
> range is -not zero- and will be stale if the mappings are not
> invalidated completely....

I can't speak to COLLAPSE_RANGE.

> Also, if page->mapping == NULL is sufficient to detect an invalidated
> page in all cases, then why does page_cache_delete() explicitly
> leave page->index intact:
> page->mapping = NULL;
> /* Leave page->index set: truncation lookup relies upon it */

Because there was, and I think still is (but might it now be xarrayed
away?), code (mainly in mm/truncate.c) which finds it convenient to
check page->index for end of range, without necessitating the overhead
of getting page lock. I've no doubt it's an (minor) optimization that
could be discarded if there were ever a need to invalidate page->index
when deleting; but nobody has required that yet.